Energy Economics 70 (2018) 26–36

Contents lists available at ScienceDirect

Energy Economics
journal homepage: www.elsevier.com/locate/eneco

Scale efficiency for multi-output cost minimizing producers: The case of

the US electricity plants夽 ,夽夽
Barnabé Walheer
International Business School Suzhou, Xi’An Jiaotong-Liverpool University, China

A R T I C L E I N F O A B S T R A C T

Article history: To know whether the optimal scale of production has been reached is valuable information for producers.
Received 20 July 2017 To date, scale efficiency measurements have only been suggested for the entire production process. For
Received in revised form 6 December 2017 multi-output producers, more detailed results are required. Hence, in this paper, we show how to pro-
Accepted 12 December 2017 vide such information at the output level. Attractively, our output-specific scale efficiency measurements
Available online 21 December 2017
are nonparametric in nature, they take the economic objective of the producers into account, they can be
defined without observing the input prices, and they are easy to interpret and to use in practice. We apply
JEL classification:
our methodology to a sample of more than 3300 US electricity plants from 1998 to 2012, producing up to
C61
10 types of electricity. We show that, while there is a scale improvement at the total electricity genera-
D24
L94 tion level, this is not the case for each of the 10 types of electricity. Also, we demonstrate that, in general,
Q29 renewable electricity presents better scale of production than non-renewable electricity. Finally, we high-
Q39 light the importance of multi-output plants in the US electricity sector, and show that this type of plant
is preferable for the production of non-renewable electricity, while single-output plants are preferable for
Keywords: renewable electricity.
Scale efficiency © 2017 Elsevier B.V. All rights reserved.
Cost minimizing
Multi-output producers
Electricity generation

1. Introduction
Assessing the optimal scale of a production process is not a
new topic in both the economic literature and the production
theory. Indeed, the concept of scale efficiency could already be
found in the works of Hanoch (1975), Panzar and Willig (1977),
Forsund and Hjalmarsson (1979), Banker (1984), Banker et al. (1984),
Färe and Grosskopf (1985), Banker and Thrall (1992), Forsund
(1996), and Golany and Yu (1997). More recent works include
those of Simar and Wilson (2002), Forsund and Hjalmarsson (2004),
Krivonozhko et al. (2004), Zelenyuk (2006, 2016), Podinovski et al.
(2009), and Peyrache (2013). These works have the investigating
of scale efficiency of the entire production process in common. Or
in other words, their methods indicate whether optimal scale is
reached for the aggregate production level. In this paper, we suggest
a technique that also provides scale efficiency results for individual
output.
Our motivation to provide output-specific scale efficiency results
is two-fold. On the one hand, by considering output-specific
indicators, the realism and the discriminatory power of the model
are naturally increased. The realism is increased since the links
between the inputs and the outputs can be modelled by allocating
the inputs to the output-specific production processes.1 The dis-
criminatory power is increased since output-specific optimization
behaviours could be assumed. On the other hand, for multi-output
producers, knowing whether the optimal scale is reached for each
output separately is clearly additional relevant information; useful
when choosing their strategy or when deciding how to allocate the
inputs.
Our scale efficiency measurements are specially designed to take
the economic objective of the producers into account. In particular,

夽 I thank the Editor-in-Chief Richard S.J. Tol, the Editor Perry Sadorsky, and the
anonymous referees for their valuable comments that substantially improved the
paper.
夽夽 This work was supported by the National Natural Science Foundation of China
under Grant No. 71750110539.
E-mail address: Barnabe.Walheer@xjtlu.edu.cn.
1
For example, employees allocated to specific output production, machines used
only to produce certain outputs. For more discussion on the allocation of inputs to
outputs, refer, for example, to Färe and Grosskopf (2000), Salerian and Chan (2005),
Despic et al. (2007), Färe et al. (2007), Tone and Tsutsui (2009), and Cherchye et al.
(2015).

https://doi.org/10.1016/j.eneco.2017.12.017
0140-9883/© 2017 Elsevier B.V. All rights reserved.
 B. Walheer / Energy Economics 70 (2018) 26–36
we assume that they are cost minimizers (the following is easily
extended to profit or revenue maximizations). Cost minimization
fits with many settings and applications, and is, by definition, a
necessary condition for profit maximization. Our model is rooted
in the nonparametric cost evaluation models initiated by Afriat
(1972), Hanoch and Rothschild (1972), Diewert and Parkan (1983)
and Varian (1984). That is, we impose very few structures on the pro-
duction process and, therefore, only the following data are required:
outputs, inputs, and input prices. The distinguishing feature of our
methodology is that by modelling each output separately, we nat-
urally give the option to assess scale efficiency at the output level.
Finally, as the observation of the input prices is rather restrictive
for some applications, we also provide alternative definitions of our
scale efficiency concepts without this assumption.
We apply our methodology to the case of US electricity plants. The
Environmental Protection Agency of the US developed a plant-level
database for 1998 to 2012. For each plant, the coal, oil, gas, nuclear,
other fossil, wind, solar, geothermal, hydro, and biomass electric-
ity generations are specified. As such, by distinguishing between
10 different types of electricity generation, this database offers a
unique opportunity to apply our methodology. In particular, the very
detailed data allow us to evaluate scale efficiency of both the indi-
vidual and aggregate electricity generation levels, and to make a
distinction between multi- and single-output producers. Therefore,
we can investigate whether multi- or single-output producers are
preferable for each of the 10 types of electricity generation. This is
valuable information for managers, regulators, and policy makers
when deciding how to allocate the production of electricity and how
to design the plants.
Moreover, our methodology offers two extra advantages in this
context. On the one hand, it gives the option to allocate the inputs to
each electricity generation type. In particular, renewable electricity is
not produced by the use of fuel, while non-renewable electricity gen-
eration requires this production factor. As such, our methodology,
which recognizes the links between production factors and electric-
ity generation, is particularly useful as it increases the realism of the
modelling of the plant production process. On the other hand, while
the data for the production factors and electricity generation are
available for the plant level, the input prices are only available at the
state level. Thus, our methodology that works with partial/without
input price data is also very attractive for that reason.
The rest of this paper unfolds as follows. Section 2 presents the
methodology. In Section 3, we apply the methodology to the case
of the US electricity plants from 1998 to 2012. Section 4 provides
conclusions.
2. Methodology
We consider that we observe producers that are cost minimizers.
In particular, we assume that they use P inputs, x ∈ RP+ , to produce Q
outputs, y ∈ RQ+ . We denote the input price vector by w ∈ RP+ . Firstly,
we assume that we observe these input prices. This will be relaxed
afterwards.
2.1. Output-specific framework
The distinguishing feature of our scale efficiency measurements
is that we make a clear distinction between aggregate and individ-
ual outputs.2 In particular, let us denote the q-th entry of y by yq .
As such, we will define scale efficiency measurements for both y and
yq . To achieve this goal, we model each output separately by its own
2
For more discussion about efficiency analysis in output-specific frameworks, refer
to Cherchye et al. (2013) for the cost setting, and Cherchye et al. (2016) for the profit
setting.
27
production process, captured by input requirement set defined as
follows for output q:
  
Iq (yq ) = xq ∈ RP+  xq can produce yq . (1)
Cost evaluation does not require us to make strong assumptions
about those sets. In fact, we follow Varian (1984) and only assume
that those sets are nested: producing less outputs cannot lead to
using more inputs.3 In this context, xq ∈ RP+ denote the inputs
used to produce the output q. In fact, those inputs are connected to
the aggregate inputs (in x). Some inputs could be used to produce
certain outputs (for example, employees, machines). That is, these
inputs are allocated to specific output production processes. Next,
some inputs could be used to produce all the outputs (for example,
infrastructure, capital), i.e. these inputs are not allocated to specific
output production processes. Formally, we have
(x)p = (x1 )p + . . . , (xQ )p , if input p is allocated, (2)
(x)p = (xq )p , if input p is not allocated. (3)
Attractively, the distinction between allocated and non-allocated
inputs provides a unifying framework that is consistent not only
with production models integrating information on the internal pro-
duction structure, but also with more standard production models
(i.e. models that do not consider allocated inputs). As a final remark,
note that the non-allocated inputs could also be interpreted as public
good (they are non-rival and non-exclusive to the output production
processes), and, therefore, they give rise to economies of scope in the
production process (see Panzar and Willig, 1981;Nehring and Puppe,
2004).
As the output-specific inputs xq could be different from the inputs
x, nothing guarantees that their price should be the same. As such,
let us denote the prices of the output-specific inputs by wq ∈ RP+ .
Note that, in general, while the input prices could be observed, the
output-specific input prices are not. The relationships between the
inputs and the output-specific inputs also imply specific relation-
ships between their prices. These prices coincide with the aggregate
prices for allocated inputs. Next, for non-allocated inputs these prices
must add up to the aggregate prices. As explained previously non-
allocated inputs could be interpreted as public good. As such, the
output-specific prices have a similar interpretation as Lindahl prices
that, by definition, sum up to the aggregate prices. In that case, the
output-specific input prices capture the economies of scope of the
production processes. Taking together, we obtain
(wq )p = (w)p , if input p is allocated, (4)

Q
(wq )p = (w)p , if input p is not allocated. (5)
q=1
As a final remark, note that the actual cost of the producers could
be rewritten exclusively by output-specific counterparts: w x =
Q q q q q
q=1 w x , where w x represents the cost of output q.
 
 
3
Iq (yq ) is nested if yq ≥ yq ⟹ Iq (yq ) ⊆ Iq yq .
28 B. Walheer / Energy Economics 70 (2018) 26–36

2.2. Cost evaluation of shadow prices). See also our discussion below when input prices
are assumed unobserved.
The starting point of the scale efficiency evaluation is the minimal
cost for each output q:
2.3. Scale efficiency
q q q q q
C (y , w, w ) = min w x . (6) Our previous definition of the technology, captured by the input
q q q x ∈I (y )
requirement sets Iq (yq ), does not assume any particular structure
Cq (yq , w, wq ) selects the minimal input vector, in the input in terms of returns-to-scale. As such, variable returns-to-scale was
requirement set Iq (yq ), to produce the output quantity yq given the implicitly assumed. To test formally for scale efficiency, we first have
 
input prices wq . C q (yq , w, wq ) ≤ wq xq , and C q (yq , w, wq ) = wq xq to define minimal costs for both the aggregate and output-specific
means that output q is produced with minimal cost, revealing cost levels under the hypothetical assumption of constant returns-to-

efficiency, while C q (yq , w, wq ) < wq xq reflects potential cost savings. scale.4 Let us denote the input requirement set satisfying the hypo-
q q
Note that, the minimal costs C (y , w, wq ) depend on the input price thetical assumption of constant returns-to-scale as Iq (yq ). It is
w, making them interdependent. This is rather intuitive, since the straightforward to define the minimal costs with respect to those
unobserved output-specific input prices depend on the input prices sets. In fact, it suffices to use Iq (yq ) instead of Iq (yq ) in Eq.(6):
(see Eqs.(4) and (5)).
Attractively, as explained before, by summing the output-specific 
C q (yq , w, wq ) = min wq xq . (8)
costs, we obtain the cost for the aggregate production level. This also xq ∈Iq (yq )
holds for the minimal costs

  The interpretation of C q (yq , w, wq ) is analogous to the interpreta-


Q
C y, w, w1 , . . . , wQ = C q (yq , w, wq ) . (7) tion of Cq (yq , w, wq ), the only difference is that the cost (in)efficient
q=1 behaviour is evaluated when assuming constant returns-to-scale.
Note also, that by definition C q (yq , w, wq ) ≤ C q (yq , w, wq ). It reflects
Clearly, the property of the costs at the aggregate output level that the input requirement set under constant returns-to-scale is,
are the same as their respective output-specific counterparts. Firstly, in general, greater than the input requirement set under variable

C q (yq , w, wq ) ≤ wq xq , implies that C(y, w, w1 , . . . , wQ ) ≤ w x, i.e. returns-to-scale, or in other words, that Iq (yq ) is included in Iq (yq ).5
at the aggregate output level, the minimal cost is bounded by the We obtain our scale efficiency index for output q and for the
actual cost. Next, if each output is produced with minimal costs, i.e. aggregate production level as follows:

C q (yq , w, wq ) = wq xq for all q, then C(y, w, w1 , . . . , wQ ) = w x, i.e.
the actual cost coincides with the minimal cost. Finally, if at least one C q (yq , w, wq )

output is produced inefficiently, i.e. C q (yq , w, wq ) < wq xq for at least SEq (yq , w, wq ) = . (9)
1 Q 
C q (yq , w, wq )
one q, we have C(y, w, w , . . . , w ) < w x.
In practice, minimal costs can be computed using linear programs.
This is particularly attractive since linear programs are easily solved. Q
  q q q
As noticed previously, the output-specific minimal costs are, by C y, w, w1 , . . . , wQ q=1 C (y , w, w )
SE y, w, w1 , . . . , wQ = 1 Q
= Q .
definition, interdependent as the unobserved output-specific input C y, w, w , . . . , w C q (yq , w, wq )
q=1
prices depend on the input prices. Attractively, we could compute (10)
all the output-specific minimal costs by solving only one program. In
fact, it suffices to evaluate the minimal costs for the aggregate out-
put level. In particular, for every producer t operating at (yt , xt ) with As C q (yq , w, wq ) ≤ C q (yq , w, wq ), SEq (yq , w, wq ) is, in general,
input price wt , the minimal cost Ct is obtained as follows (LP-1): smaller than 1. A value of one indicates scale efficiency behaviour.
When SEq (yq , w, wq ) < 1, it reveals scale inefficiency, which could be
due to decreasing or increasing returns-to-scale.6 As for the output-

Q
Ct = max
q
Ct specific value, SE(y, w, w1 , . . . , wQ ) = 1 reflects scale efficiency, while
Q
Ct1 ,...,Ct ∈R+ q=1 a value smaller than one implies more scale inefficiency behaviour.
Q Q
w1t ,...,wt ∈R+ Attractively, in that case, the source(s) of inefficiency could be found
s.t. ∀q ∈ {1, . . . , Q}, the following holds: simply by looking at the values of the SEq (yq , w, wq ). As a final remark,
q q q q q note that our definition of scale efficiency at the aggregate produc-
(C − 1) : Ct ≤ wt xs for all s : ys ≥ yt ,
  tion level is coherent with the one of Färe and Grosskopf (1985). The
q
(C − 2) : wt = (wt )p for p an allocated input, only difference is that we base our measurement on output-specific
p
technologies.
Q 
 
q The linear program for the minimal costs under the hypothetical
(C − 3) : wt = (wt )p for p a non-allocated input.
p assumption of constant returns-to-scale has a structure that is for-
q=1
mally analogous to the one of (LP-1). In particular, for every producer
q t operating at (yt , xt ) with input price wt , the minimal cost under the
In words, (C-1) picks, for every output q, the minimal cost Ct
when comparing the evaluated producer t to the dominating pro-
q
ducers (i.e. those that produce more outputs than yt ). Note that
it is why we have to impose that the input requirement sets are
0 : x ∈ I (y ) ⟹ kx ∈
nested, otherwise we can only compare producer t to producers that (y ) satisfies constant returns-to-scale if ∀k ∈ R+
4 Iq q q q q q

produce exactly the same output quantity. (C-2) and (C-3) make
sure that the unknown output-specific input prices are correctly
specified (see Eqs.(4) and (5)). As a final remark, it could seem
counter-intuitive to maximize a cost function. In fact, the maximiza-
tion selects the most favourable output-specific input prices (notion
Iq (kyq ).
(y ) is directly related to Iq (yq ), since Iq (yq ) = k(xq ) ∈ Iq (kyq ), ∀k > 0 .
5 Iq q
6
In practice, it is enough to evaluate the minimal cost when assuming non-
increasing returns-to-scale and compare to Cq (yq , w, wq ). If they are equal, scale
inefficiency is due to decreasing returns-to-scale. Otherwise, scale inefficiency is due
to increasing returns-to-scale.
 B. Walheer / Energy Economics 70 (2018) 26–36 29

hypothetical assumption of constant returns-to-scale is obtained as As the computed input prices are the most favourable, the mini-
follows (LP-2): mal cost C y, w1 , . . . , wQ and C y, w1 , . . . , wQ provide upper bounds
for the minimal costs when the input prices are assumed observed.

Q
Ct = max
q
Ct Importantly, since no input prices are available, the constraints on
Q
Ct1 ,...,Ct ∈R+ q=1 the output-specific input prices in Eqs.(4) and (5) are irrelevant. The
Q Q
w1t ,...,wt ∈R+ alternative definitions of scale efficiency are obtained by plugging-in
s.t. ∀q ∈ {1, . . . , Q}, the following holds: the new definitions of the minimal costs (12) and (13) in Eqs.(9) and
    (10). Attractively, we can also evaluate those minimal costs by linear
q q q q q q q
(C − 1) : Ct ≤ wt fs xs for all s : fs ys ≥ yt , programs. For every producer t operating at (yt , xt ), the minimal cost
 
q
(C − 2) : wt = (wt )p for p an allocated input, Ct is obtained by solving (LP-3):
p
Q 
 
q 
Q
q
(C − 3) : wt = (wt )p for p a non-allocated input, Ct = max Ct
p Q
q=1 Ct1 ,...,Ct ∈R+ q=1
  1 Q
wt ,...,wt ∈R+
Q
q q q
(C − 4) : fs = inf f ∈ R+
0 | fys ≥ yt .
s.t. ∀q ∈ {1, . . . , Q},
the following holds:
   
q q
q q q q q
(LP-2) is very similar to (LP-1) except that the input-output of the (C − 1) : ≤
Ct fs xs for all s : fs ys ≥ yt ,
wt
q
dominating producers are rescaled by the factor fs . In fact, this factor  
q q q
(C − 2) : fs = inf f ∈ R+ 0 | fys ≥ yt .
is present exactly to take the hypothetical assumption of constant
returns-to-scale into account in the linear program. This is captured wt xt = 1.
q
by the restriction on the set of the possible value of fs . Note that,
q
(LP-2) coincides with (LP-1) if fs = 1, ∀q, ∀s, i.e. no rescaling of the As explained previously, when input prices are not observed, no
input-output.7 We end this part by one important remark: there is constraints (except that they have to be non-negative) are put on the
an interesting relationship between scale efficiency at both levels.8 output-specific input prices. A common way to proceed, that dates
In fact, scale efficiency at the aggregate level could be obtained as from the work of Charnes et al. (1978), is to normalize the actual
a weighted sum of the scale efficiencies at the output-specific level cost to unity (i.e. wt xt = 1). The normalization is used to make the
where the weights are the output-specific minimal cost shares: computed minimal costs comparable, i.e. the benchmark value is 1.9
The minimal costs Ct are obtained, for every producer t, by setting
q
fs = 1, ∀q, ∀s in (LP-3). Finally, note that extra constraints on the
  Q computed prices could easily be added in (LP-3). For example, lower
C q (yq , w, wq )
SE y, w, w1 , . . . , wQ = Q SEq (yq , wq , wq ) . and upper bounds could be included, as it is the case in our Applica-
C q (yq , w, wq )
q=1 q=1 tion where we use state-level input prices as lower and upper bounds
(11) for the unknown plant-level input prices (See Section 3 for more
details). Those extra constraints are added to increase the realism of
2.4. Input price availability the computed prices.

As it is defined, our aggregate and output-specific scale efficiency 3. Application

indicators depend on the observation of the input price vectors (w).
In the following we relax this assumption. We believe that it is par- We apply our methodology to the case of the US electric-
ticularly attractive since we can keep the advantage of basing our ity plants. Investigating the scale optimality, building on a cost
indicators on economic objective without facing the disadvantage minimization behaviour, of electricity plants has already been
of observing the input prices. As done before for the output-specific considered by several authors. See, for example, Christensen
input prices, we choose the input prices that maximize the minimal and Greene (1976), Nelson (1985), Nelson and Wohar (1985),
costs (i.e. the shadow prices). As such, we evaluate the producers Krautmann and Solow (1988), Nemoto et al. (1993), Burns
in the best possible way in the absence of true input price infor- and Weyman-Jones (1996), Filippini (1996), Considine (2000),
mation, which gives the benefit of the doubt to the producers. We Filippini and Wild (2001), Kleit and Terrell (2001), Maloney
obtain (2001), Rhine (2001), Hiebert (2002), Fraquelli et al. (2005),
⎧ ⎫ Kopsakangas-Savolainen and Svento (2008), Akkemik (2009), Arcos
  
Q ⎨Q ⎬ and De Toledo (2009), Assaf et al. (2011), Kumbhakar et al. (2015),
1 Q q q q q q q
C y, w , . . . , w = C (y , w ) = max C (y , w, w ) , Ajayi et al. (2017).10 The cost minimizing framework is mostly cho-
Q ⎩ ⎭
q=1 w∈R q=1
+ sen in this context as the output side of the production process is
(12) rather exogenous to the plant (fixed, in a sense, by the demand),
⎧ ⎫ while the plants can still control the input side given the electricity
  
Q ⎨Q ⎬
generation. As such, the input side is rather endogenous to the plants.
C y, w1 , . . . , wQ = C q (yq , wq ) = max C q (yq , w, wq ) .
w∈R+
Q ⎩ ⎭
q=1 q=1
(13)

7
Note that other returns-to-scale assumptions are easily implemented by replac-
ing R+0 by (0, 1], [1, ∞) for the decreasing or increasing returns to scale assumption,
respectively. For more details, see also Petersen (1990) and Bogetoft (1996) in a techni-
cal nonparametric setting. Also, note that for (C-4), we assume that, if it is not possible
q q
to find such a fs (i.e. the solution is the empty set), then fs = +∞, i.e. no constraints
are put on the minimal costs.
8
Zelenyuk (2016) obtains a similar result in a group aggregation context based on
revenue maximization.
9
Note that any value could, in principle, be used for the normalization, 1 is more
convenient and commonly used in this case.
10
Note that in several works other methods are used to compute the cost functions
(such as a stochastic frontier model). In principle, these alternative methods could be
extended to include output-specific indicators. The advantages of the nonparamet-
ric model are its easy use and that no strong assumptions are required about the
production process. See our discussion of Eq. (1).
30 B. Walheer / Energy Economics 70 (2018) 26–36

This implies that cost minimization is preferable to profit or revenue

Fig. 1. Production process of the electricity plants.
maximization in this context.
The distinguishing feature of our methodology is to consider each
type of electricity separately, instead of modelling only the aggregate
electricity generation. As such, our analysis, while remaining consis-
tent with the previous works, offers the advantage to provide more
detailed results, without making extra assumptions on any aspect of
the production process. Moreover, our methodology offers two extra
advantages in this context. One, it allows us to allocate the inputs to
each electricity generation type, by explicitly recognizing the links
between production factors and electricity generation. In particular,
it is clear that renewable electricity is not produced by the use of fuel,
while non-renewable electricity generation is. Two, our methodol-
ogy also works with partial input price data. This is very attractive
in this context as, while the data for the production factors and elec-
tricity generation are available for the plant-level, the input prices
are only available at the state-level.
We tackle two important questions in this empirical part. Firstly,
we compute scale efficiency of 10 types of electricity generations:
coal, oil, gas, nuclear, other fossil, wind, solar, geothermal, hydro,
and biomass. This is attractive since it is not obvious that the scale
performances are the same for each type of electricity production.
Moreover, our sample consists of more than 3300 plants over a large
period (1998–2012), meaning that our results are trustworthy. Next,
we investigate whether multi- or single-output producers are prefer-
able for each of the 10 types of electricity generation. We believe
that the answers to these two questions are valuable information
for managers, regulators, and policy makers when deciding how to
allocate the production of electricity and how to design the plants.
To present our empirical application, we first define the produc-
tion process of the plants. Next, we present our data and highlight
the importance of multi-output producers in the US. Afterwards, we
present the scale efficiency results. Finally, we provide the shadow
prices for the inputs.

Table 1
Descriptive statistics (MWh) and percentage of electricity generation per type.

Year Total electricity Non-renewable electricity Renewable electricity

Average Median Max Average Median Max Average Median Max Percentage

2012 932,991 54,907 31,933,916 1,421,533 72,088 31,933,916 201,069 38,620 26,461,174 11.99%
2010 997,874 61,727 31,199,935 1,545,660 82,487 31,199,935 184,443 40,978 18,351,775 10.12%
2009 972,270 56,840 30,661,851 1,494,580 74,293 30,661,851 192,098 40,642 20,987,193 10.91%
2007 1,054,755 58,641 26,782,391 1,648,342 105,251 26,782,391 177,627 36,096 21,632,495 9.12%
2005 1,057,669 66,996 25,807,446 1,643,632 121,436 25,807,446 191,635 41,163 20,474,048 9.84%
2004 1,043,974 62,891 28,112,609 1,625,071 102,214 28,112,609 191,788 37,513 18,917,123 10.01%
2000 1,083,721 77,744 30,380,571 1,646,675 133,821 30,380,571 207,732 44,034 21,765,745 10.35%
1999 1,063,401 81,071 30,415,572 1,590,314 136,600 30,415,572 234,004 46,131 24,967,234 12.04%
1998 1,041,813 81,273 30,301,045 1,574,153 144,480 30,301,045 233,544 47,031 21,038,360 12.27%

Year Coal Oil Gas Nuclear Other fossil Hydro Biomass Wind Solar Geothermal

2012 2,697,422 23,062 392,136 11,743,938 221,226 134,013 54,078 41,027 428,174 86,535
48.24% 0.84% 15.06% 24.67% 0.39% 8.71% 1.50% 0.09% 0.02% 0.38%
2010 3,136,804 32,581 327,569 12,174,188 200,186 128,759 54,144 41,391 425,636 79,218
54.14% 1.14% 11.62% 23.47% 0.35% 7.36% 1.36% 0.09% 0.02% 0.35%
2009 2,968,009 33,543 312,489 12,161,972 211,640 124,572 55,482 41,114 424,032 77,486
53.09% 1.23% 11.37% 24.05% 0.35% 8.00% 1.37% 0.09% 0.02% 0.36%
2007 3,394,674 51,684 328,110 12,139,194 188,655 132,720 62,607 46,482 428,685 103,422
56.18% 1.74% 10.99% 22.13% 0.51% 6.59% 1.32% 0.09% 0.02% 0.34%
2005 3,383,420 94,413 303,306 11,761,769 208,463 130,399 52,782 45,437 440,625 119,388
55.79% 3.28% 9.98% 21.37% 0.55% 7.24% 1.28% 0.08% 0.02% 0.34%
2004 3,312,531 92,455 294,057 11,873,207 205,628 140,100 60,177 51,926 444,484 101,018
55.53% 3.22% 9.71% 21.89% 0.47% 7.23% 1.41% 0.09% 0.02% 0.35%
2000 3,233,258 74,824 376,365 11,554,690 215,519 140,456 64,908 57,448 480,860 94,111
53.62% 2.77% 13.07% 20.66% 0.50% 7.37% 1.33% 0.09% 0.02% 0.36%
1999 3,094,710 83,327 361,269 11,203,248 248,562 142,549 58,681 56,591 476,607 94,328
52.45% 3.09% 12.71% 20.44% 0.56% 8.68% 1.37% 0.08% 0.02% 0.39%
1998 3,133,357 92,052 352,481 10,365,985 252,520 134,758 51,457 56,211 487,747 91,118
53.38% 3.41% 12.45% 19.32% 0.50% 8.99% 1.25% 0.08% 0.02% 0.39%
 B. Walheer / Energy Economics 70 (2018) 26–36 31

Table 2 the Environmental Protection Agency provides price data for the
Production factors. states. As such, we use those prices to increase the realism of the
Year Boilers Generators Nameplate capacity Fuel computed prices. In particular, we use the plant-level prices as lower
(number) (number) (MW) (MMBtu) and upper bounds for the unknown plant-level prices. The goal of
2012 1.15 3.68 247 10,488,370 these bounds is to avoid trivial and/or unrealistic prices (as too close
2010 1.17 3.40 246 11,607,583 to zero or too large). See our discussion at the end of Section 2 for
2009 1.16 3.38 245 11,103,202 more details. Next, as the data for the number of boilers and gener-
2007 0.65 3.38 241 12,593,262
ators are also provided by the eGRID system, we could incorporate
2005 1.00 3.38 241 12,795,503
2004 0.99 3.42 241 12,894,084 them as extra inputs. This is done in, for example, Sarkis and Cordeiro
2000 0.61 2.75 230 13,474,571 (2012). In that case, P = 4. As the results are very similar and for the
1999 0.59 2.73 230 13,237,346 sake of simplicity, we present the case with two inputs, i.e. P = 2.
1998 0.58 2.06 229 13,249,417

3.1. Data and plant production process 3.2. Descriptive statistics and importance of the multi-output plants

We use data of the eGRID system developed by the Environmental
Protection Agency in the US. In particular, we use all the databases
available between 1998 and 2012 (there are no databases available
after 2012). Unfortunately, the databases are not provided for each
year but for nine years: 2012, 2010, 2009, 2007, 2005, 2004, 2000,
1999, and 1998. There are also two more databases in 1996 and 1997,
but the input labels are different; thus, we do not take those two
extra databases into account. This results in a sample of 3389 plants.
As explained in the Introduction, the first distinguishing feature
of our empirical analysis is that we model each type of electricity
separately. Each plant can produce up to 10 types of electricity: coal,
oil, gas, nuclear, other fossil, wind, solar, geothermal, hydro, and
biomass (in fact, the maximum is 10 and the average is 1.40 on the
period). We have Q = 10. As done by the eGRID system, we regroup
the ten types of electricity into two categories: renewable (wind,
solar, geothermal, hydro, and biomass) and non-renewable (coal,
oil, gas, nuclear, other fossil). By summing the ten types of electric-
ity generation, we obtain the total electricity generation. Therefore,
we obtain three levels of comparison: ten types of electricity gen-
eration (coal, oil, gas, nuclear, other fossil, wind, solar, geothermal,
hydro, and biomass), two aggregate categories (renewable and non-
renewable), and the total generation.
While the output side of the production process is described in
much detail by the eGRID system, the input side is not so detailed.
Indeed, only the fuel consumption is provided by eGRID system. For-
tunately, we could proxy the missing inputs (such as infrastructure,
labour) by the nameplate capacity, provided by the eGRID system.
This strategy has already been used, for example, by Tone and Tsut-
sui (2011), Sarkis and Cordeiro (2012), Cherchye et al. (2015), and
Walheer (2017, 2018) in a similar context. We have two inputs, i.e.
P = 2. An attractive feature of our technique, as discussed previ-
ously, is that it gives the option to allocate the inputs to the outputs.
In this context, this is particularly relevant as the fuel input is clearly
not used to produce the renewable electricity. We summarize the
production process of a typical plant in Fig 1.
We end this part by two remarks. Firstly, no data for the input
prices are reported for the plants by the eGRID system. Nevertheless,
We present the descriptive statistics in Table 1. The first part
of Table 1 (top) gives the averages, medians, and maxima for the
total, non-renewable, and renewable electricity generation for the
period. Clearly, plants produce more non-renewable than renewable
electricity for the period. This is confirmed by both the averages
and the medians. Moreover, as highlighted by the maxima, the
largest plants are those producing non-renewable electricity. Next,
renewable electricity generation represents around 10% of the total
electricity production, and this percentage is more or less constant
for the period. The second part of Table 1 (bottom) presents the
relative importance of each type of electricity. It reveals that coal
electricity generation represents around 50% of the total electricity
generation for the period, while nuclear electricity generation repre-
sents around 25%, gas electricity generation represents around 10%,
and hydro electricity generation representsslightly less than 10%. For
non-renewable electricity, the coal and oil electricity generation are
decreasing, while the gas and nuclear are increasing. For renewable
electricity, they are all more or less stable for the period.
Table 2 presents the descriptive statistics for the input side of the
production process. While the fuel consumption slowly decreases for
the period, the nameplate capacity increases. It means that plants are
becoming bigger. This is also confirmed by the rise of the number of
generators and boilers.
As explained in the Introduction, the second distinguishing fea-
ture of our empirical analysis is that we differentiate between multi-
and single-output plants. To justify the importance of this distinction
in this context, we present in Table 3, the proportion of multi-output
producers for the total, non-renewable, and renewable electricity
generation, and the percentage of electricity produced and produc-
tion factors used by the multi-output plants.
While the percentage of multi-output producers decreases from
almost 40% to 30% over the period, they are important as they pro-
duce around 60% of the total electricity, and represent 60% of the
nameplate capacity and 90% of the fuel consumption. This impor-
tance is higher for non-renewable electricity generation as more than
50% of the producers are multi-output producers, and they repre-
sent around 70% of the production of this type of electricity. Both

Table 3
Importance of the multi-output producers.

Year Total Non-renewable Renewable Production factors

Multi Prod Multi Prod Multi Prod Namplate capacity Fuel

2012 31% 60% 54% 66% 16% 12% 59% 90%

2010 33% 64% 56% 70% 17% 13% 61% 92%
2009 32% 63% 55% 68% 17% 12% 60% 92%
2007 34% 67% 57% 72% 16% 14% 64% 93%
2005 35% 68% 60% 73% 16% 12% 65% 94%
2004 34% 67% 59% 72% 16% 12% 63% 92%
2000 39% 68% 66% 73% 18% 13% 67% 92%
1999 38% 67% 65% 73% 18% 12% 65% 92%
1998 37% 67% 64% 74% 16% 11% 65% 92%
32 B. Walheer / Energy Economics 70 (2018) 26–36

Table 4
Averages, medians, and Kolmogorov-Smirnov p-values.

Year Total Non-renewable Renewable

All Multi Single All Multi Single All Multi Single

2012 0.58 0.50 0.62 0.49 0.48 0.51 0.67 0.57 0.69
2010 0.61 0.57 0.64 0.51 0.54 0.48 0.72 0.69 0.72
2009 0.53 0.54 0.53 0.47 0.51 0.41 0.61 0.68 0.59
2007 0.46 0.56 0.40 0.54 0.54 0.54 0.37 0.62 0.32
2005 0.60 0.60 0.60 0.57 0.57 0.57 0.63 0.70 0.61
2004 0.55 0.57 0.54 0.52 0.53 0.51 0.58 0.70 0.56
2000 0.58 0.61 0.55 0.58 0.58 0.56 0.58 0.71 0.55
1999 0.51 0.58 0.47 0.57 0.57 0.57 0.45 0.58 0.42
1998 0.52 0.58 0.48 0.54 0.56 0.51 0.49 0.66 0.46
2012 0.62 0.53 0.69 0.48 0.51 0.46 0.71 0.60 0.73
2010 0.70 0.64 0.72 0.53 0.58 0.46 0.76 0.76 0.76
2009 0.57 0.61 0.56 0.49 0.55 0.37 0.62 0.76 0.60
2007 0.46 0.60 0.36 0.54 0.55 0.53 0.28 0.71 0.22
2005 0.63 0.66 0.61 0.62 0.63 0.59 0.63 0.78 0.62
2004 0.57 0.64 0.55 0.55 0.58 0.48 0.57 0.77 0.55
2000 0.58 0.70 0.52 0.64 0.66 0.56 0.54 0.77 0.51
1999 0.51 0.65 0.41 0.63 0.64 0.58 0.40 0.64 0.35
1998 0.51 0.66 0.39 0.63 0.63 0.58 0.39 0.73 0.38
2012 > 1998 0.00 0.01 0.00 0.00 0.02 0.00 0.00 0.99 0.00
2012 > 2010 0.00 0.01 0.00 0.00 0.00 0.00 1.00 0.97 1.00
2010 > 2009 0.00 0.00 0.00 0.00 0.01 0.00 0.00 0.61 0.00
2009 > 2007 0.00 0.19 0.00 0.34 0.17 0.90 0.00 0.00 0.00
2007 > 2005 0.94 0.03 0.84 0.00 0.03 0.00 0.82 0.99 0.83
2005 > 2004 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.45 0.00
2004 > 2000 0.00 1.00 0.00 0.96 1.00 0.17 0.00 0.20 0.00
2000 > 1999 0.00 0.00 0.00 0.00 0.00 0.15 0.00 0.00 0.00
1999 > 1998 0.00 0.00 0.00 0.00 0.00 0.00 0.25 0.89 0.21

these numbers decrease over the period. For the renewable electric-
ity production, the multi-output producers represent around 16% of
the number of producers, and generate around 12% of the electricity.
Both these numbers are stable over the period.
In Table 6 in the Appendix, we give the number of producers for
each of the 10 types of electricity generation, and the proportion of
multi-output producers. A first observation is that the proportion of
multi-output producers is very high for the coal and solar electric-
ity generation, as they represent more than 90% of the producers.
Next, they represent around 70% of the producers for the oil, gas, and
biomass electricity generation. The majority of producers of nuclear,
hydro, and wind electricity generation are single-output produc-
ers, while there are no multi-output producers for the geothermal
electricity generation. Finally, note that the proportion of multi-
output producers is slowly decreasing for every type of electricity
generation, except for solar electricity generation.
for every plant. We also add extra constraints on the prices as dis-
cussed before. We use three tools to present our results without losing
too much information. Indeed, we have the scale efficiency scores for
the 3389 plants for the 10 outputs and the nine years. In particular,
an obvious choice is to give the averages and the medians for each
year. While these two statistics give a good first approximation of
the change of scale efficiency, relying only on the averages and medi-
ans to conclude if the scale efficiency has improved or not could be
restrictive in this context. Namely, because of the size of our sample.
As such, a more formal way to check the existence of an improve-
ment is to make use of the two-sample Kolmogorov-Smirnov test (KS
test). This is a nonparametric test that checks whether the distribu-
tions of two samples are equal or not. In our context, we calibrate the
test to check whether an improvement exists, i.e. if the distribution
has moved to the right. As such, we also provide the p-values of the
two-sample KS test. If the p-value is smaller than 5%, we can reject

3.3. Scale efficiency results Table 6

Shadow prices for nameplate capacity and fuel.

The previous descriptive analysis has set the stage by revealing Year All Non-renewable Renewable
the main characteristics of the US production plants, and the impor- Multi Single Multi Single
tance of the multi-output plants. In this part, we present the results
2012 6.51 5.52 5.8 6.81 6.83
of the scale efficiency scores. Those scores are computed using (LP-3)
2010 7.73 6.65 6.81 8.01 7.86
2009 8.08 7.52 7.65 8.56 8.45
Table 5 2007 8.09 7.52 7.64 8.56 8.45
Kolmogorov-Smirnov p-values. 2005 6.41 6.35 6.21 6.99 6.84
2004 5.42 5.31 5.22 5.68 5.76
Year Renewable > Non-renewable Multi > single
2001 5.51 5.45 5.62 5.62 5.71
1999 5.5 5.31 5.54 5.63 5.64
All Multi Single Total Non-renewable Renewable
1998 5.81 5.22 5.32 6.01 5.98
2012 0.00 0.00 0.00 0.00 0.00 0.00 2012 3.09 3.01 3.12 – –
2010 0.00 0.00 0.00 0.00 0.32 0.07 2010 3.85 3.84 3.86 – –
2009 0.00 0.00 0.00 0.09 0.36 0.20 2009 3.86 3.7 4.01 – –
2007 0.00 0.00 0.00 0.14 0.02 0.74 2007 4.63 4.4 4.71 – –
2005 0.00 0.00 0.00 0.00 0.01 0.13 2005 5.63 5.56 5.65 – –
2004 0.00 0.00 0.00 0.00 0.00 0.32 2004 4.45 4.4 4.48 – –
2000 0.00 0.00 0.00 0.00 0.00 0.38 2004 5.45 5.32 4.98 – –
1999 0.00 0.00 0.00 0.15 0.00 0.26 1999 5.64 5.61 5.67 – –
1998 0.00 0.00 0.00 0.06 0.00 0.24 1998 5.21 5.19 5.23 – –
 B. Walheer / Energy Economics 70 (2018) 26–36
the null hypothesis, and conclude that the distribution has moved to
the right implying an improvement of scale efficiency. The results for
the total, non-renewable, and renewable electricity generations are
shown in Table 4. For each type of electricity generation, we consider
three cases: all the producers, the multi-output producers, and the
single-output producers.
Firstly, we can conclude that there is an improvement of scale
efficiency over the period for the total electricity generation. Except
for 2007, the averages and medians have increased, and the p-values
of the KS test are close to zero. The same conclusion holds if we take
only the multi- or the single-output producers into account. Note
that the averages have decreased for the multi-output producers, but
the p-values of the KS test are close to zero. This reveals the use-
fulness of this test in our context. Next, non-renewable electricity
generation shows a slight improvement, while the improvement on
renewable electricity generation is rather high. This is confirmed by
the p-values of the KS test. Again, similar results hold when looking
only at the multi- and single-output producers. Next, two important
rankings seem present in Table 4. One, renewable electricity gener-
ation seems to present higher scale efficiency than non-renewable
electricity. Two, single-output producers seem less efficient on non-
renewable electricity, but more efficient on renewable than multi-
output producers. The differences are more pronounced for renew-
able electricity generation. To formally check those two observations,
we again make use of the KS test. The results are available in Table 5.
Our first observation about the relationship between renewable
and non-renewable electricity generation is clearly confirmed, as the
KS test p-values are all close to zero, implying that the distribution
of renewable electricity is always larger than that for non-renewable
electricity. Our second observation is also confirmed. Indeed, the
majority of the KS test p-values are close to zero for the total and non-
renewable electricity, but not for the renewable electricity. As such,
multi-output producers are justified for non-renewable electricity,
while single-output producers are preferred for renewable electricity.
Finally, we present in Tables 7–9 (available in the Appendix),
the averages, medians, and KS test p-values of the scale efficiency
scores for each type of electricity generation. We give below the
main findings. Firstly, the averages and medians for the nuclear
generation are the highest, and, as shown by the KS test p-values,
they are increasing. The performance of the three other types of
non-renewable electricity generation (i.e. coal, oil, and gas) are
decreasing; this explains the relative worse performances for non-
renewable electricity generation discussed previously. Then, there
are high scores in hydro electricity generation (and increasing),
biomass electricity generation (and decreasing), and wind electricity
generation (and increasing), which explains the better performances
for renewable electricity generation highlighted before.
When comparing single- and multi-output producers, it is clear
that single-output producers perform better for hydro, biomass, and
wind electricity generation, which explains the better performances
of this type of producer for renewable electricity. The performances
of the single-output plant for geothermal electricity generation are
also high, but slowly decreasing. Note that no results are presented
for geothermal electricity generation for the multi-output plants as
only single-output plants produce that type of electricity. Interest-
ingly, the performances of the multi-output producers are higher
for solar electricity generation. For non-renewable electricity, the
multi-output producers perform better for nuclear, coal, gas, and oil
electricity generation; but as discussed previously, the differences
are more pronounced for renewable generation.11
11
Those results are also confirmed by KS tests. For the sake of compactness (i.e. 10
types of electricity for nine years), we do not present those results in the paper, but
they are available by request from the author.
33
3.4. Shadow prices
We end our application by providing the shadow prices. As
explained previously, one of the advantages of the suggested
methodology for scale efficiency is that it works when prices are not
or partially available. For our US plant context, this is attractive as
only the state-level input prices are provided by the Environmen-
tal Protection Agency of the US. We provide in Table 6 the average
shadow prices when assuming variable returns-to-scale. Note that
similar results could be obtained when assuming constant returns-
to-scale, but since the latter assumption is more restrictive, we rely
on the former. This also reveals that assumptions about the technol-
ogy should be well motivated as they have an impact on the results.
As done previously for the scale efficiency results, we make a distinc-
tion between renewable and non-renewable electricity generation,
and between single- and multi-output producers. The shadow prices
for nameplate capacity are given in dollar per MW, and for fuel in
dollar per MMBtu.
As explained in detail in Section 2, the shadow prices represent
the most favourable prices in the absence of price information. As
such, they are not the estimators of the true unknown prices, and
have thus to be interpreted carefully. Nevertheless, they not only
provide interesting useful information for the managers of the plants,
but also for policy-makers and regulators. At this juncture, we point
out that, in general, estimating the production cost of the plants is
not an easy task. In the US, the Environmental Protection Agency
is putting much effort into estimating these costs.12 The provided
shadow prices could help in that task.
An initial observation is that the price of nameplate capacity for
the producers of renewable electricity is, on average, higher than the
price for the producers of non-renewable electricity. This could be
explained as it is the only input used by the producers of renewable
electricity. Next, on average, the prices of the single-output producers
are higher for both inputs than those of the multi-output produc-
ers. This might be explained by the presence of economies of scope.
Indeed, a reason for producing more than one type of electricity is to
obtain a decrease of the average total cost (see also our discussion of
Eq.(3)). The smaller input prices might be interpreted in that sense.
4. Conclusion
In this paper, we have shown how to provide scale efficiency
indicators at the output-level. This is particularly attractive for multi-
output producers since it represents valuable information. Moreover,
our output-specific scale efficiency measurements are nonparamet-
ric in nature, they take the economic objective of the producers into
account, they can be defined without observing the input prices, and
they are easy to interpret and to use in practice.
We have applied our methodology to the case of the US electricity
plants. Using the plant-level database developed by the Environmen-
tal Protection Agency of the US, we evaluated the scale efficiency
of more than 3300 US electricity plants from 1998 to 2012 for 10
different types of electricity generation (coal, oil, gas, nuclear, other
fossil, wind, solar, geothermal, hydro, and biomass). We show that,
while there is a scale improvement at the total electricity generation
level, this is not the case for each of the 10 types of electricity. Also,
we demonstrated that, in general, renewable electricity presents
better scale of production than non-renewable electricity. Next, we
highlighted the importance of the multi-output plants in the US
electricity sector, and show that those plants are preferable for the
production of non-renewable electricity, while single-output plants
are preferable for renewable electricity.
12
See the following website for more detail: www.eia.gov/outlooks/capitalcost.
34 B. Walheer / Energy Economics 70 (2018) 26–36

Appendix A

Table 7
Number of producers and percentage of multi-output producers.

Year Coal Oil Gas Nuclear Other fossil Hydro Biomass Wind Solar Geothermal

2012 538 1032 1144 63 136 1178 339 53 18 27

92% 73% 63% 8% 99% 2% 69% 9% 94% 0%
2010 581 1095 1167 65 151 1213 357 57 16 28
93% 75% 65% 9% 99% 2% 69% 11% 94% 0%
2009 586 1075 1149 65 148 1219 358 55 15 28
92% 74% 64% 6% 97% 3% 69% 7% 88% 0%
2007 588 1133 1171 65 177 1223 354 53 12 28
91% 75% 67% 11% 98% 3% 69% 2% 92% 0%
2005 589 1196 1160 65 166 1219 351 54 12 28
92% 76% 70% 11% 98% 3% 68% 4% 92% 0%
2004 591 1174 1142 65 165 1214 355 52 11 28
92% 76% 69% 11% 97% 3% 66% 2% 91% 0%
2000 602 1330 1258 65 192 1225 345 51 11 27
95% 79% 76% 14% 94% 3% 74% 0% 91% 0%
1999 602 1299 1252 65 213 1221 342 51 11 29
95% 78% 74% 14% 94% 3% 74% 0% 91% 0%
1998 593 1264 1228 64 191 1219 323 51 11 28
95% 78% 73% 14% 94% 1% 72% 0% 91% 0%

Table 8
Averages, medians, and Kolmogorov-Smirnov p-values: all producers.

Year Coal Oil Gas Nuclear Other fossil Hydro Biomass Wind Solar Geothermal

2012 0.57 0.33 0.45 0.89 0.42 0.69 0.60 0.67 0.47 0.70
2010 0.74 0.35 0.37 0.89 0.45 0.72 0.73 0.62 0.54 0.78
2009 0.71 0.24 0.37 0.84 0.50 0.56 0.75 0.67 0.55 0.64
2007 0.67 0.38 0.40 0.87 0.54 0.26 0.70 0.65 0.61 0.66
2005 0.75 0.42 0.41 0.89 0.51 0.58 0.77 0.71 0.59 0.67
2004 0.71 0.35 0.41 0.91 0.48 0.52 0.77 0.68 0.60 0.72
2000 0.77 0.37 0.47 0.84 0.54 0.51 0.77 0.65 0.43 0.62
1999 0.74 0.37 0.45 0.85 0.50 0.37 0.64 0.68 0.48 0.63
1998 0.66 0.26 0.56 0.81 0.53 0.42 0.70 0.65 0.47 0.70
2012 0.59 0.33 0.42 0.90 0.42 0.74 0.61 0.63 0.46 0.74
2010 0.76 0.33 0.32 0.91 0.44 0.76 0.78 0.60 0.43 0.78
2009 0.74 0.19 0.34 0.83 0.48 0.57 0.78 0.68 0.50 0.64
2007 0.72 0.38 0.37 0.87 0.52 0.17 0.73 0.64 0.71 0.64
2005 0.78 0.45 0.42 0.90 0.53 0.60 0.81 0.72 0.72 0.69
2004 0.74 0.33 0.38 0.92 0.48 0.54 0.81 0.70 0.60 0.72
2000 0.80 0.34 0.51 0.90 0.62 0.48 0.82 0.71 0.42 0.66
1999 0.76 0.36 0.45 0.91 0.58 0.33 0.68 0.70 0.46 0.69
1998 0.70 0.20 0.66 0.87 0.57 0.37 0.75 0.70 0.38 0.78
2012 > 1998 0.97 0.00 0.00 0.00 0.31 0.00 0.83 0.71 0.27 0.61
2012 > 2010 1.00 0.02 0.00 0.79 0.44 0.99 0.94 0.13 0.77 0.96
2010 > 2009 0.00 0.00 0.05 0.00 0.63 0.00 0.98 0.91 0.74 0.00
2009 > 2007 0.00 0.97 0.98 0.86 0.89 0.00 0.00 0.20 0.60 0.54
2007 > 2005 1.00 0.89 0.00 0.94 0.02 0.65 0.99 0.93 0.43 0.86
2005 > 2004 0.00 0.00 0.08 0.98 0.09 0.00 0.38 0.40 0.80 0.86
2004 > 2000 0.88 0.59 0.28 0.11 0.41 0.00 0.16 0.56 0.13 0.32
2000 > 1999 0.00 0.00 0.00 0.12 0.15 0.00 0.00 0.69 0.86 0.69
1999 > 1998 0.00 0.00 0.15 0.00 0.59 0.23 0.68 0.39 0.51 0.98

Table 9
Averages, medians, and Kolmogorov-Smirnov p-values: multi-output producers.

Year Coal Oil Gas Nuclear Other fossil Hydro Biomass Wind Solar Geothermal

2012 0.59 0.41 0.52 0.89 1.00 0.51 0.58 0.40 0.49 –
2010 0.81 0.40 0.43 0.91 0.55 0.57 0.70 0.47 0.54 –
2009 0.77 0.20 0.43 0.84 0.86 0.40 0.72 0.25 0.55 –
2007 0.68 0.52 0.48 0.87 0.77 0.14 0.68 0.14 0.61 –
2005 0.79 0.54 0.51 0.89 0.66 0.47 0.74 0.44 0.59 –
2004 0.75 0.38 0.51 0.91 0.73 0.40 0.73 0.65 0.61 –
2000 0.82 0.39 0.61 0.91 0.79 0.35 0.74 – 0.43 –
1999 0.77 0.49 0.56 0.91 0.56 0.30 0.60 – 0.48 –
1998 0.61 0.23 0.68 0.87 0.76 0.24 0.68 – 0.49 –
2012 0.61 0.37 0.49 0.90 1.00 0.57 0.60 0.42 0.46 –
2010 0.85 0.40 0.43 0.91 0.55 0.70 0.76 0.54 0.42 –
2009 0.80 0.10 0.41 0.83 0.86 0.49 0.77 0.20 0.56 –
2007 0.73 0.51 0.46 0.87 0.75 0.07 0.72 0.14 0.71 –
2005 0.83 0.54 0.61 0.91 0.66 0.55 0.80 0.44 0.72 –
(continued on next page)
 B. Walheer / Energy Economics 70 (2018) 26–36 35

Table 9 (continued)

Year Coal Oil Gas Nuclear Other fossil Hydro Biomass Wind Solar Geothermal

2004 0.78 0.33 0.55 0.92 0.77 0.50 0.79 0.65 0.67 –
2000 0.83 0.32 0.67 0.92 0.83 0.41 0.78 – 0.43 –
1999 0.80 0.50 0.60 0.92 0.49 0.29 0.65 – 0.47 –
1998 0.63 0.12 0.77 0.88 0.84 0.21 0.73 – 0.54 –
2012 > 1998 0.87 0.00 0.00 0.00 0.06 0.01 0.97 – 0.61 –
2012 > 2010 0.97 0.00 0.00 0.98 0.06 0.84 0.93 0.78 0.77 –
2010 > 2009 0.13 0.00 0.05 0.00 0.85 0.00 0.93 0.13 0.72 –
2009 > 2007 0.00 0.67 0.64 0.93 0.06 0.00 0.00 1.00 0.59 –
2007 > 2005 0.98 0.02 0.00 0.93 0.07 1.00 0.98 0.14 0.43 –
2005 > 2004 0.06 0.00 0.05 0.98 0.37 0.06 0.40 0.56 0.80 –
2004 > 2000 0.77 0.00 0.10 0.10 0.38 0.31 0.14 – 0.37 –
2000 > 1999 0.01 0.01 0.00 0.15 0.15 0.07 0.00 – 0.93 –
1999 > 1998 0.00 0.00 0.67 0.00 0.49 0.96 0.88 – 0.48 –

Table 10
Averages, medians, and Kolmogorov-Smirnov p-values: single-output producers.

Year Coal Oil Gas Nuclear Other fossil Hydro Biomass Wind Solar Geothermal

2012 0.57 0.31 0.42 0.88 0.41 0.69 0.63 0.69 0.04 0.70
2010 0.73 0.33 0.34 0.71 0.45 0.72 0.78 0.64 0.45 0.78
2009 0.70 0.25 0.33 0.84 0.49 0.57 0.81 0.71 0.50 0.64
2007 0.67 0.33 0.37 0.84 0.54 0.26 0.75 0.66 0.54 0.66
2005 0.75 0.38 0.37 0.87 0.51 0.59 0.84 0.72 0.55 0.67
2004 0.71 0.34 0.36 0.89 0.47 0.52 0.84 0.68 0.43 0.72
2000 0.77 0.36 0.42 0.53 0.53 0.51 0.85 0.71 0.41 0.74
1999 0.74 0.34 0.41 0.61 0.50 0.38 0.74 0.73 0.39 0.72
1998 0.66 0.27 0.51 0.58 0.51 0.42 0.75 0.71 0.28 0.80
2012 0.59 0.30 0.36 0.89 0.42 0.74 0.65 0.64 0.04 0.74
2010 0.76 0.31 0.29 0.83 0.44 0.76 0.82 0.61 0.45 0.78
2009 0.74 0.23 0.31 0.83 0.48 0.57 0.82 0.70 0.50 0.64
2007 0.71 0.32 0.34 0.87 0.52 0.17 0.76 0.65 0.54 0.64
2005 0.78 0.39 0.35 0.89 0.53 0.60 0.84 0.72 0.55 0.69
2004 0.73 0.33 0.31 0.93 0.47 0.54 0.83 0.70 0.43 0.72
2000 0.79 0.35 0.45 0.82 0.62 0.48 0.88 0.72 0.41 0.69
1999 0.76 0.32 0.38 0.86 0.59 0.33 0.75 0.73 0.39 0.70
1998 0.70 0.23 0.62 0.80 0.55 0.37 0.82 0.71 0.28 0.80
2012 > 1998 0.99 0.00 0.01 0.05 0.30 0.00 0.42 0.67 0.23 0.87
2012 > 2010 1.00 0.99 0.00 0.18 0.57 0.99 0.93 0.06 1.00 0.96
2010 > 2009 0.00 0.00 0.03 0.44 0.53 0.00 0.55 0.98 1.00 0.00
2009 > 2007 0.00 1.00 0.99 0.05 0.92 0.00 0.00 0.13 1.00 0.54
2007 > 2005 1.00 0.97 0.00 0.50 0.04 0.64 0.96 0.93 0.56 0.86
2005 > 2004 0.00 0.00 0.32 0.84 0.10 0.00 0.71 0.38 0.14 0.86
2004 > 2000 0.91 0.90 0.63 0.32 0.29 0.00 0.33 0.74 0.08 0.56
2000 > 1999 0.00 0.03 0.06 0.73 0.22 0.00 0.00 0.60 0.14 0.64
1999 > 1998 0.00 0.00 0.17 0.25 0.39 0.22 0.67 0.36 0.14 0.98

Appendix B. Supplementary data
Supplementary data to this article can be found online at https://
doi.org/10.1016/j.eneco.2017.12.017.
References
Afriat, S., 1972. Efficiency estimation of production functions. Int. Econ. Rev. 13,
568–598.
Akkemik, K.A., 2009. Cost function estimates, scale economies and technological
progress in the Turkish electricity generation sector. Energy Policy 1, 204–213.
Arcos, A., De Toledo, P.A., 2009. An analysis of the Spanish electrical utility industry:
economies of scale, technological progress and efficiency. Energy Econ. 31 (3),
473–481.
Ajayi, V., Weyman-Jones, T., Glass, A., 2017. Cost efficiency and electricity market
structure: a case study of OECD countries. Energy Econ. 65, 283–291.
Assaf, A.G., Barros, C.P., Managi, S., 2011. Cost efficiency of Japanese steam power gen-
eration companies: a Bayesian comparison of random and fixed frontier models.
Appl. Energy 88, 1441–1446.
Banker, R.D., 1984. Estimating most productive scale size using data envelopment
analysis. Eur. J. Oper. Res. 17, 35–44.
Banker, R.D., Charnes, A., Cooper, W.W., 1984. Some models for estimating techni-
cal and scale inefficiencies in data envelopment analysis. Manag. Sci. 30,
1078–1092.
Banker, R.D., Thrall, R.M., 1992. Estimation of returns to scale using data envelopment
analysis. Eur. J. Oper. Res. 62, 74–84.
Bogetoft, P., 1996. DEA on relaxed convexity assumptions. Manag. Sci. 42, 457–465.
Burns, P., Weyman-Jones, T.G., 1996. Cost functions and cost efficiency in electricity
distribution: a stochastic frontier approach. Bull. Econ. Res. 48 (1), 41–64.
Charnes, A., Cooper, W.W., Rhodes, E., 1978. Measuring the efficiency of decision
making units. Eur. J. Oper. Res. 2, 429–444.
Cherchye, L., De Rock, B., Dierynck, B., Roodhooft, F., Sabbe, J., 2013. Opening the black
box of efficiency measurement: input allocation in multi-output settings. Oper.
Res. 61, 1148–1165.
Cherchye, L., De Rock, B., Walheer, B., 2015. Multi-output efficiency with good and bad
outputs. Eur. J. Oper. Res. 240, 872–881.
Cherchye, L., De Rock, B., Walheer, B., 2016. Multi-output profit efficiency and direc-
tional distance functions. Omega 61 (C), 100–109.
Christensen, L.R., Greene, W.H., 1976. Economies of scale in US electric power
generation. J. Polit. Econ. 655–676.
Considine, T.J., 2000. Cost structures for fossil fuel-fired electric power generation.
Energy J. 83–104.
Despic, O., Despic, M., Paradi, J., 2007. DEA-r: ratio-based comparative efficiency
model, its mathematical relation to DEA and its use in applications. J. Prod. Anal.
28, 33–44.
Diewert, W.E., Parkan, C., 1983. Linear programming tests of regularity conditions for
production frontiers. Quant. Stud. Prod. Prices.
Färe, R., Grosskopf, S., 2000. Network DEA. Socioecon. Plann. Sci. 34, 35–49.
Färe, R., Grosskopf, S., 1985. A nonparametric cost approach to scale efficiency. Scand.
J. Econ. 87, 594–604.
Färe, R., Grosskopf, S., Whittaker, G., 2007. Network DEA. In: Zhu, J., Cook, W. (Eds.),
Modeling Data Irregularities and Structural Complexities in Data Envelopment
Analysis. Springer.
Filippini, M., 1996. Economies of scale and utilization in the Swiss electric power
36 B. Walheer / Energy Economics 70 (2018) 26–36
distribution industry. Appl. Econ. 28 (5), 543–550.
Filippini, M., Wild, J., 2001. Regional differences in electricity distribution costs and
their consequences for yardstick regulation of access prices. Energy Econ. 23 (4),
477–488.
Forsund, F.R., 1996. On the calculation of the scale elasticity in DEA models. J. Prod.
Anal. 7, 283–302.
Forsund, F.R., Hjalmarsson, L., 1979. Generalised Farrell measures of efficiency: an
application to milk processing in Swedish dairy plants. Econ. J. 89, 294–315.
Forsund, F.R., Hjalmarsson, L., 2004. Calculating scale elasticity in DEA models. J. Oper.
Res. Soc. 55, 1023–1038.
Fraquelli, G., Piacenza, M., Vannoni, D., 2005. Cost savings from generation and
distribution with an application to Italian electric utilities. J. Regul. Econ. 28 (3),
289–308.
Golany, B., Yu, G., 1997. Estimating returns to scale in DEA. Eur. J. Oper. Res. 103,
28–37.
Hanoch, G., 1975. The elasticity of scale and the shape of average costs. Am. Econ. Rev.
65, 492–497.
Hanoch, G., Rothschild, M., 1972. Testing assumptions of production theory: a non-
parametric approach. J. Polit. Econ. 80, 256–275.
Hiebert, L.D., 2002. The determinants of the cost efficiency of electric generating
plants: a stochastic frontier approach. South. Econ. J. 935–946.
Kleit, A.N., Terrell, D., 2001. Measuring potential efficiency gains from deregulation of
electricity generation: a Bayesian approach. Rev. Econ. Stat. 83 (3), 523–530.
Kopsakangas-Savolainen, M., Svento, R., 2008. Estimation of cost-effectiveness of the
Finnish electricity distribution utilities. Energy Econ. 30 (2), 212–229.
Krautmann, A.C., Solow, J.L., 1988. Economies of scale in nuclear power generation.
South. Econ. J. 70–85.
Krivonozhko, V.E., Utkin, O.B., Volodin, A.V., Sablin, I.A., Patrin, M., 2004. Constructions
of economic functions and calculations of marginal rates in DEA using parametric
optimization methods. J. Oper. Res. Soc. 55, 1049–1058.
Kumbhakar, S.C., Amundsveen, R., Kvile, H.M., Lien, G., 2015. Scale economies, tech-
nical change and efficiency in Norwegian electricity distribution, 1998-2010. J.
Prod. Anal. 43 (3), 295–305.
Maloney, M.T., 2001. Economies and diseconomies: estimating electricity cost
functions. Rev. Ind. Organ. 19 (2), 165–180.
Nehring, K., Puppe, C., 2004. Modelling cost complementarities in terms of joint
production. J. Econ. Theory 118, 252–264.
Nelson, R.A., 1985. Returns to scale from variable and total cost functions: evidence
from the electric power industry. Econ. Lett. 18 (2), 271–276.
Nelson, R.A., Wohar, M.E., 1985. Regulation, scale economies, and productivity in
steam electric generation. Int. Econ. Rev. 57–79.
Nemoto, J., Nakanishi, Y., Madono, S., 1993. Scale economies and over-capitalization in
Japanese electric utilities. Int. Econ. Rev. 431–440.
Panzar, J.C., Willig, R.D., 1977. Economies of scale in multi-output production. Q. J.
Econ. 91, 481–493.
Panzar, J.C., Willig, R.D., 1981. Economies of scope. Am. Econ. Rev. 71 (2), 268–272.
Petersen, N.C., 1990. Data envelopment analysis on a relaxed set of assumptions.
Manag. Sci. 36, 305–314.
Peyrache, A., 2013. Industry structural inefficiency and potential gains from mergers
and break-ups: a comprehensive approach. Eur. J. Oper. Res. 230, 422–430.
Podinovski, V.V., Forsund, F.R., Krivonozhko, V., 2009. A simple derivation of scale
elasticities in data envelopment analysis. Eur. J. Oper. Res. 197, 149–153.
Rhine, R., 2001. Economies of scale and optimal capital in nuclear and fossil fuel
electricity production. Atl. Econ. J. 29 (2), 203–214.
Sarkis, J., Cordeiro, J.J., 2012. Ecological modernization in the electrical utility industry:
An application of a bads-goods DEA model of ecological and technical efficiency.
Eur. J. Oper. Res. 219, 386–395.
Salerian, J., Chan, C., 2005. Restricting multiple-output multiple-input DEA models by
disaggregating the output-input vector. J. Prod. Anal. 24, 5–29.
Simar, L., Wilson, P.W., 2002. Non-parametric tests of returns to scale. Eur. J. Oper. Res.
139, 115–132.
Tone, K., Tsutsui, M., 2009. Network DEA: a slacks-based measure approach. Eur. J.
Oper. Res. 197, 243–252.
Tone, K., Tsutsui, M., 2011. Applying an Efficiency Measure of Desirable and Undesir-
able Outputs in DEA to U.S. Electric Utilities approach. J. CENTRUM Cathedra 4,
236–249.
Varian, H.R., 1984. The non-parametric approach to production analysis. Econometrica
52, 579–598.
Walheer, B., 2017. Cost Malmquist productivity index: an output-specific approach for
group comparison. J. Prod. Anal.Accepted.
Walheer, B., 2018. Disaggregation of the cost Malmquist productivity index with joint
and output-specific inputs. Omega 75, 1–12.
Zelenyuk, V., 2006. A scale elasticity measure for directional distance function and its
dual: theory and DEA estimation. Eur. J. Oper. Res. 228, 592–600.
Zelenyuk, V., 2016. Aggregation of scale efficiency. Eur. J. Oper. Res. 240, 269–277.